Active Sampling for Knowledge Discovery from Biomedical Data

1

• Active Sampling for Knowledge Discovery from
  Biomedical Data
• Sriharsha Veeramachaneni, ITC-IRST
• Francesca Demichelis, ITC-IRST, Harvard Medical
  School
• Emanuele Olivetti, ITC-IRST
• Paolo Avesani, ITC-IRST

ECML/PKDD 2005, Porto, Portugal
2
Outline

• Biomarker Evaluation for Cancer Characterization
• Costs Involved
• Formal Definition of Cost-Constrained Biomarker
  Evaluation
• Outline of Solution
• Experimental Results
• Conclusions

3
Cancer Characterization

• Development of cancer diagnostic/prognostic
  models from Data.
• clinical features (e.g. status of the patient,
  relapse)
• histological features (given by the pathologist)
• grade of disease
• tumor dimension
• number of lymphonodes
• ...
• Molecular biomarker features
• protein expression by immunohistochemistry
• gene amplification by fluorescence in situ
  hybridization

4
Molecular biomarker features

• Biomarkers are molecular reagents used to test
  biological samples,
• that add information to the clinical and
  pathology data (improve predictive model for
  diagnosis/prognosis).
• They should impact clinical care.

5
Tissue Micro Array (TMA)

• TMA block is a paraffin block with 100-1000
  tissue cores in array fasion
• each core is 0.6mm Ø from regions of interest of
  the donor block
• each TMA block can be sliced in 100-500 thin
  slides
• each slide tested independently
  (immunohistochemistry, florescent hybridization)

6
Donor Blocks
TMA block
TMA Slides
7
Biomarker Evaluation for Clinical Prediction
Process

• Collect biological samples from monitored
  patients along with features (clinical,
  histological)
• Set up an initial dataset. Each record is
• C Clinical information (eg. Dead/Alive, Relapse)
• X Histological features
• Prepare TMA blocks from the biological samples.
• add biomarkers and measure outcomes Y (intensity,
  extension)
• Analyze the ability of the biomarker Y to predict
  C given X.
• NOTE This is a one shot process (i.e. one
  experiment).

8
Cost Issues

• Limitations in clinical analysis of tissues
• limited availability of tissues
• cumbersome and time consuming procedure
• cost of biochemical reagents

9
Active Sampling for Feature Evaluation
Data set
Choose the most informative (C,X) to sample for Y
No
Sample at (C, X)
Evaluate predictive ability of Y
Budget Exhausted?
Yes
Output estimate of feature utility
10
Active Sampling for Biomarker Evaluation

• collect known features from monitored patients
  (clinical, histological)
• set up an initial dataset (C clinical , X
  histological)
• create an initial model to predict C given
  already known X
• select intelligently TMA block/slide to improve
  current model and pay for that information
• add biomarkers and measure outcomes (Y)
• build a new model to predict class labels C given
  X,Y features.
• iterate, going back to (4) until budget exhausted
• Predict error rate of the classifier that uses
  both X and Y
• NOTE This is an iterative process.

11
Active Sampling Choosing the most informative
sample

• We want to estimate the error-rate (e) of the
  classifier using both X and Y accurately.
• We should probe where we think sampling will
  improve the estimate of the error-rate the most.
• We do this by minimizing the expected MSE after
  sampling for each (C, X) and choose the lowest.
• We show that the above criterion is equivalent to
  maximizing the expected difference in the
  estimates before and after sampling at (C, X).

12
Active Sampling Issues

• Finding a heuristic to approximate the
  theoretically optimal sampling strategy.
• Dealing with the bias of the sampling scheme.
• Computational considerations.

13
Experiments Breast Cancer Dataset Description

• 2 alternative class labels (clinical data)
• C1 status of the patient (dead/alive)
• C2 presence/absence of tumor relapse
• 3 known features (histological data)
• X1 tumor type
• X2 pathologist's evaluation of lymphonodes
• X3 pathologist's evaluation of morphology
• 4 biomarkers features
• Y1 nuclei expressing ER
• Y2 nuclei expressing PGR
• Y3 score colour intensity stained area by
  P53
• Y4 score colour intensity stained area by
  cerbB
• Size of dataset 400 record with missing values
  (see later)
• All features reduced to binary values based on
  domain conventions.

14
Experiments Description

• Class Label (C) Known Features (X)
  New Feature (Y) Size ()
• I dead/alive all histological
  information PGR 160
• II dead/alive all histological
  information P53 164
• III dead/alive all histological
  information ER 152
• IV dead/alive all histological
  information cerbB 170
• V relapse all histological
  information PGR 157
• VI relapse all histological
  information P53 161
• VII relapse all histological
  information ER 149
• VIII relapse all histological
  information cerbB 167
• IX dead/alive PGR, P53, ER
  cerbB 196
• X relapse PGR, P53, ER
  cerbB 198

15
Some Results
16
Conclusions

• We presented a preliminary solution for
  cost-constrained biomarker evaluation
• For the same accuracy in the prediction of
  feature relevance we need fewer samples.
• For the application, this means that we need to
  use up less of the valuable biological tissue
  resource.
• Allows testing more biomarkers on the same amount
  of tissue samples.

17
Future Work

• Extend the algorithm to
• Other classifiers
• Continuous features
• Batch active learning, where we sample more than
  one sample at a time
• Study how much a myopic strategy for sampling is
  worse than non-myopic.

18
Thank You